Deep learning model as an inversion tool for resonant ultrasound spectroscopy of piezoelectric materials
Wuyi Yang, Shanshan Sun, Jing Hu, Liguo Tang, Lei Qin, Zhenglin Li, Wenyu Luo
Abstract
Device fabrication based on piezoelectric materials requires prior characterization of full matrix constants. For this, the Institute of Electrical and Electronics Engineers standard on piezoelectricity suggests the use of ultrasonic pulse-echo and electric resonance methods. However, these techniques tend to provide inconsistent characterization, because they require multiple samples with drastically different sizes. Resonant ultrasound spectroscopy (RUS) is a promising alternative, because it uses only a single sample for characterization, thus ensuring self-consistent results. The inverse problem of finding material constants from resonant frequencies is often solved using the nonlinear least squares method despite its being a time-consuming algorithm. Herein, deep learning (DL) is introduced in the inversion procedure of RUS. After the DL network is trained, the material constants are determined with high efficiency. The practicability and reliability of the combination of DL and RUS are demonstrated by characterizing the full tensor constants of LiNbO3 single crystals.